“A compute cluster used to train a frontier AI model can typically support the simultaneous operation of 100,000 instances of that same model for inference.”